A Gradient Boosting Regression Model for Early-Stage Screening of Total Biogas Yield from Organic Substrates

Abstract

Early-stage estimation of total biogas yield is useful for preliminary screening of anaerobic digestion substrates when detailed kinetic parameters, long-term reactor data and substrate-specific biochemical methane potential tests are not yet available. This study develops a compact Gradient Boosting Regression model for estimating total biogas yield from basic substrate and operating descriptors commonly reported in anaerobic digestion studies. A literature-informed reconstructed benchmark dataset containing 120 records from twelve organic substrate groups was used. The response variable was total biogas yield, expressed as mL g−1 volatile solids (VS) added; methane yield and methane content were not used as interchangeable target variables. Seven input variables were considered: volatile solids content, total solids content, carbon-to-nitrogen ratio, residence time descriptor, digestion temperature, pH and organic loading rate. The model was evaluated using the coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), five-fold cross-validation and residual diagnostics. On the test subset, the model achieved R2 = 0.9885, MAE = 13.40 mL g−1 VS, RMSE = 16.16 mL g−1 VS and MAPE = 3.73%. The mean five-fold cross-validation score was R2 = 0.9720 ± 0.0180, indicating stable interpolation performance within the investigated data domain. Feature-importance analysis identified C/N ratio as the dominant predictor, followed by volatile solids and total solids. However, residual diagnostics showed heteroscedasticity, and the low apparent importance of pH and temperature was attributed to their narrow ranges in the compiled dataset, rather than to biological irrelevance. The proposed model is therefore intended as an early-stage screening tool for total biogas yield estimation, not as a universal methane yield or design-level prediction model.

More by Zafar Turakulov

• Reaction-Center-Guided Reranking of Transformer Product Predictions with a Product-Conditioned Graph Neural NetworkResearch Square · 2026
• A PROCESS-AWARE DATA PREPROCESSING AND TIME-DELAY IDENTIFICATION FRAMEWORK FOR CEMENT CLINKER ENERGY MONITORINGZenodo · 2026
• Campus Decarbonization in Central Asia Through a Whole-System Sustainability Transition: A Case Study of the Tashkent Institute of Chemical Technology2026
• Techno-Economic Comparison of Carbon Capture Technologies with Exhaust Gas Recirculation in NGCC Power Plants2026